The reproduction record

Twelve days were selected by even spacing across the full archive span (2024-02-01 → 2026-06-15) — not cherry-picked — and rebuilt from the raw L1 FITS, through the frozen parsers and builders, then compared against the frozen dataset.

Two distinct claims were measured separately, because they are not the same thing:

• Content identity — the values are identical. This is the scientific claim: a rebuild must reproduce the same measurements.
• Byte identity — the parquet file bytes are identical. Stronger, but partly a property of the writer environment (parquet embeds library metadata), so it was measured rather than assumed.

Day	T1 content	T1 byte	T2 content	Rebuild
2024-02-01	✅	✅	✅	0.6 s
2024-03-10	✅	✅	✅	0.4 s
2024-04-20	✅	✅	✅	0.4 s
2024-05-31	✅	✅	✅	0.4 s
2025-08-09	✅	✅	✅	0.4 s
2025-09-30	✅	✅	✅	0.4 s
2025-11-12	✅	✅	✅	0.4 s
2025-12-23	✅	✅	✅	0.4 s
2026-02-01	✅	✅	✅	0.4 s
2026-03-15	✅	✅	✅	0.4 s
2026-05-04	✅	✅	✅	0.4 s
2026-06-15	✅	✅	✅	0.4 s

Content identity: T1 12/12, T2 12/12. Byte identity: T1 12/12, T2 12/12.

Both the 340-channel spectral arrays (T2) and the scalar light-curve table (T1) reproduce exactly. Row counts match (1,440 rows/day) on every sample.

The pipeline contains no non-deterministic operation by construction:

Property	Guarantee
No randomness	No RNG anywhere in parsers or builders (no sampling, no shuffling, no initialisation)
No imputation	NaN is preserved as the missing-data sentinel; never filled, never dropped (verified by AST scan + tests)
No reordering	The parser preserves archive order losslessly; HK sorting is an explicit out-of-parser utility
Order-independent resolution	The Version Resolution Engine builds a coverage map in two phases (collect → resolve); shuffled candidate order yields identical maps, proven on real data
Order-independent aggregation	All minute statistics are sum/mean/max/min over finite values
Fixed hashing	Provenance SHAs come from the frozen Phase 0.5.1 manifest, not recomputed at build time

Version-resolution stability across three independent full builds: 1,065,572 owned (minute, detector) pairs, 48,604 conflicts, R1 47,328 / R2 1,276 / R3 0 / F-14 0 — byte-stable every time.

T7 provenance_manifest contains parsed_at_utc, a build wall-clock timestamp. This column differs between builds by design — it records when a build ran. Every other T7 column is deterministic.

Consequently:

• The reproducible content hash of T7 excludes parsed_at_utc (NON_REPRODUCIBLE_COLUMNS in freeze_dataset.py).
• The file SHA-256 of T7 in the manifest pins this frozen artifact exactly, as it does for every other file.

This is stated rather than engineered away: removing the timestamp would have modified the dataset after freezing, and the timestamp is legitimate provenance.

Field	Value
Python	3.12.12
numpy / pandas	1.26.4 / 2.2.2
pyarrow / astropy	16.1.0 / 6.1.7
Platform	macOS-26.5.2-arm64 (Apple Silicon)
Lockfile SHA-256	6899e001b1c4d64a5aec01b3ee1d2cfce2a3ff38ab64f827a75ab9d55fd13d3f

Scope of the byte-identity claim: demonstrated in this pinned environment. Byte identity depends on the parquet writer (pyarrow==16.1.0); a different pyarrow may produce different bytes for identical data. Content identity is the portable guarantee; byte identity is the observed, stronger result here. Both are reported so the distinction is not lost.

Stage	Time
Per-day rebuild (T1 + T2, from raw FITS)	0.4 s median
Full archive rebuild (424 SoLEXS days + 389 HEL1OS orbits)	93.75 min
Archive extraction (827 ZIPs → 135 GB)	6.8 min
Scientific validation	~20 min
Sample verification (this report)	~6 s

See DATASET_MANIFEST.md §5. In short: install from the lockfile, run extract_archive.py, run build_canonical.py, then freeze_dataset.py and compare dataset_hash against 43fd0e228b28ae6bc7e468c3acf68722768bd62b73798eb6631e9e6233b71ed9.

Verdict: the dataset is reproducible from the raw archive. Content-identical by construction and byte-identical in the pinned environment, with one documented, deliberate exception (parsed_at_utc).